A Probabilistic Novelty Detection Methodology Based on the Order-Frequency Spectral Coherence

Abstract

The purpose of this paper is to develop a methodology that utilises the order-frequency spectral coherence to detect novel (i.e. unobserved) second-order cyclostationary components. In the novelty detection methodology, a probabilistic model of the healthy data is utilised to detect, localise and trend novelties in the form of damage that manifest as second-order cyclostationary components in vibration signals. The methodology is unique in the sense that on the one hand, the spectral properties are retained and multiple harmonics of the fault frequency component are used during the condition inference process; while on the other hand, the methodology is simple and efficient to implement. A numerical gearbox model is used to generate vibration signals and to simulate bearing and distributed gear damage. The methodology is applied to the simulated vibration signals, generated under varying speed conditions, which demonstrates very promising results.